An Offline Image Auditing System for Legacy Meter Reading Systems in Developing Countries: A Machine Learning Approach

Abstract

The developing countries are challenged with overbilling and underbilling, due to manual meter reading, which results in consumer dissatisfaction and loss of revenue. The existing automated meter reading (AMR) solutions are expensive; hence, sample-based manual snap auditing systems are introduced to control such meter reading inaccuracies. In these systems, the meter reader, besides reading, also collects meter images, which are used to manually audit the meter’s accuracy. Although such systems are inexpensive, they are limited in their ability to be sustainable and ensure 100% accurate meter readings. In this paper, a novel offline optical character recognition (OCR) system-based Snap Audit system is proposed and tested for its efficient and real-time 100% accurate meter reading capabilities. The experimental results on 5,000 real-world instances show that the proposed approach processes an image in 0.05 seconds with 94% accuracy. Moreover, the developed approach is evaluated with four state-of-the-art algorithms: region convolution neural network (RCNN), nanonets, Fast-OCR, and PyTesseract. The results provide evidence that our new system design along with novel approach is more robust and efficient as compared to existing algorithms by 43.6%.